diff --git a/paddle/fluid/operators/reduce_ops/reduce_op.h b/paddle/fluid/operators/reduce_ops/reduce_op.h
index 65cca94814e88111239aef3559285d6fe321a72d..ff1ddb4175feff0b059c7ba43e77b306197a11a0 100644
--- a/paddle/fluid/operators/reduce_ops/reduce_op.h
+++ b/paddle/fluid/operators/reduce_ops/reduce_op.h
@@ -265,67 +265,6 @@ class ReduceKernel : public framework::OpKernel<T> {
         framework::TransToPhiDataType(cast_out_dtype), output);
   }
 };
-template <typename DeviceContext, typename OutT, typename Functor>
-class BoolReduceKernel : public framework::OpKernel<OutT> {
- public:
-  void Compute(const framework::ExecutionContext& context) const override {
-    bool reduce_all = context.Attr<bool>("reduce_all");
-    auto* input = context.Input<Tensor>("X");
-    auto* output = context.Output<Tensor>("Out");
-    output->mutable_data<OutT>(context.GetPlace());
-
-    auto dims = context.Attr<std::vector<int>>("dim");
-    bool keep_dim = context.Attr<bool>("keep_dim");
-
-    // The dims has full dim, set the reduce_all is True
-    const auto& input_dim_size = context.Input<Tensor>("X")->dims().size();
-    std::set<int> dims_set(dims.begin(), dims.end());
-    bool full_dim = true;
-    for (auto i = 0; i < input_dim_size; i++) {
-      if (dims_set.find(i) == dims_set.end()) {
-        full_dim = false;
-        break;
-      }
-    }
-    reduce_all = (reduce_all || full_dim);
-
-    if (reduce_all) {
-      // Flatten and reduce 1-D tensor
-      auto x = EigenVector<OutT>::Flatten(*input);
-      auto out = EigenScalar<OutT>::From(*output);
-      auto& place =
-          *context.template device_context<DeviceContext>().eigen_device();
-      auto reduce_dim = Eigen::array<int, 1>({{0}});
-      Functor functor;
-      functor(place, &x, &out, reduce_dim);
-    } else {
-      int ndim = input->dims().size();
-      int rdim = dims.size();
-      // comments for accelerating compiling temporarily.
-      if (ndim > 6) {
-        HandleLargeDim<DeviceContext, OutT, Functor>(context, input, output,
-                                                     dims, keep_dim);
-      } else {
-        HANDLE_DIM(6, 5);
-        HANDLE_DIM(6, 4);
-        HANDLE_DIM(6, 3);
-        HANDLE_DIM(6, 2);
-        HANDLE_DIM(6, 1);
-        HANDLE_DIM(5, 4);
-        HANDLE_DIM(5, 3);
-        HANDLE_DIM(5, 2);
-        HANDLE_DIM(5, 1);
-        HANDLE_DIM(4, 3);
-        HANDLE_DIM(4, 2);
-        HANDLE_DIM(4, 1);
-        HANDLE_DIM(3, 2);
-        HANDLE_DIM(3, 1);
-        HANDLE_DIM(2, 1);
-        HANDLE_DIM(1, 1);
-      }
-    }
-  }
-};
 
 template <typename DeviceContext, typename T, typename Functor>
 void LaunchReduceGradKernel(const framework::ExecutionContext& context,
diff --git a/paddle/phi/kernels/cpu/reduce_grad_kernel.cc b/paddle/phi/kernels/cpu/reduce_grad_kernel.cc
index 78a7ae8d415b5d4b18fdf8e469576db50f739e38..4b3b1fc16e9c426daccb6364f926a7c44ab63e59 100644
--- a/paddle/phi/kernels/cpu/reduce_grad_kernel.cc
+++ b/paddle/phi/kernels/cpu/reduce_grad_kernel.cc
@@ -99,8 +99,8 @@ void ReduceSumGradKernel(const Context& dev_ctx,
 
   ReduceGradKernel<Context, T, funcs::SumGradFunctor, true>(dev_ctx,
                                                             x,
-                                                            out_grad,
                                                             paddle::none,
+                                                            out_grad,
                                                             dims,
                                                             keep_dim,
                                                             reduce_all,
@@ -121,8 +121,8 @@ void ReduceMeanGradKernel(const Context& dev_ctx,
                           DenseTensor* x_grad) {
   ReduceGradKernel<Context, T, funcs::MeanGradFunctor, true>(dev_ctx,
                                                              x,
-                                                             out_grad,
                                                              paddle::none,
+                                                             out_grad,
                                                              dims,
                                                              keep_dim,
                                                              reduce_all,
diff --git a/paddle/phi/kernels/impl/frobenius_norm_grad_kernel_impl.h b/paddle/phi/kernels/impl/frobenius_norm_grad_kernel_impl.h
index 65d903a7fe426c6eed6cba6f38e8c636001d47b0..1877a4ecc227e36cad474d8a5c11de5ff7a1946c 100644
--- a/paddle/phi/kernels/impl/frobenius_norm_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/frobenius_norm_grad_kernel_impl.h
@@ -33,7 +33,7 @@ void FrobeniusNormGradKernel(const Context& ctx,
                              DataType out_dtype,
                              DenseTensor* dx) {
   ReduceGradKernel<Context, T, funcs::FrobeniusNormGradFunctor>(
-      ctx, x, dout, out, axis, keep_dim, reduce_all, in_dtype, out_dtype, dx);
+      ctx, x, out, dout, axis, keep_dim, reduce_all, in_dtype, out_dtype, dx);
 }
 
 }  // namespace phi
diff --git a/paddle/phi/kernels/impl/reduce_grad.h b/paddle/phi/kernels/impl/reduce_grad.h
index f56d3d3ed50f7e72910115f7ec28914a5eade2e8..0b1c43b5f040227b691fca6bdc257fbf8c697957 100644
--- a/paddle/phi/kernels/impl/reduce_grad.h
+++ b/paddle/phi/kernels/impl/reduce_grad.h
@@ -87,8 +87,8 @@ template <typename Context,
           bool kNoNeedBufferY = false>
 void ReduceGradKernel(const Context& dev_ctx,
                       const DenseTensor& x,
-                      const DenseTensor& out_grad,
                       const paddle::optional<DenseTensor>& out,
+                      const DenseTensor& out_grad,
                       const std::vector<int64_t>& dims,
                       bool keep_dim,
                       bool reduce_all,
diff --git a/paddle/phi/kernels/impl/reduce_max_grad_kernel_impl.h b/paddle/phi/kernels/impl/reduce_max_grad_kernel_impl.h
index 4a74416e3916492e6d3a40e09ca347db485fff7c..0a0c1abac808646db157b39ef929b007a25e6985 100644
--- a/paddle/phi/kernels/impl/reduce_max_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/reduce_max_grad_kernel_impl.h
@@ -24,8 +24,8 @@ namespace phi {
 template <typename T, typename Context>
 void ReduceMaxGradKernel(const Context& dev_ctx,
                          const DenseTensor& x,
-                         const DenseTensor& out_grad,
                          const DenseTensor& out,
+                         const DenseTensor& out_grad,
                          const std::vector<int64_t>& dims,
                          bool keep_dim,
                          bool reduce_all,
@@ -34,8 +34,8 @@ void ReduceMaxGradKernel(const Context& dev_ctx,
                          DenseTensor* x_grad) {
   ReduceGradKernel<Context, T, funcs::MaxOrMinGradFunctor>(dev_ctx,
                                                            x,
-                                                           out_grad,
                                                            out,
+                                                           out_grad,
                                                            dims,
                                                            keep_dim,
                                                            reduce_all,
diff --git a/paddle/phi/kernels/impl/reduce_min_grad_kernel_impl.h b/paddle/phi/kernels/impl/reduce_min_grad_kernel_impl.h
index baaa544f137366f1e0343c25bc373cc08350f7fd..965fc686e27838ccc3ec73b7ae7d8c5267cd59bf 100644
--- a/paddle/phi/kernels/impl/reduce_min_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/reduce_min_grad_kernel_impl.h
@@ -24,8 +24,8 @@ namespace phi {
 template <typename T, typename Context>
 void ReduceMinGradKernel(const Context& dev_ctx,
                          const DenseTensor& x,
-                         const DenseTensor& out_grad,
                          const DenseTensor& out,
+                         const DenseTensor& out_grad,
                          const std::vector<int64_t>& dims,
                          bool keep_dim,
                          bool reduce_all,
@@ -34,8 +34,8 @@ void ReduceMinGradKernel(const Context& dev_ctx,
                          DenseTensor* x_grad) {
   ReduceGradKernel<Context, T, funcs::MaxOrMinGradFunctor>(dev_ctx,
                                                            x,
-                                                           out_grad,
                                                            out,
+                                                           out_grad,
                                                            dims,
                                                            keep_dim,
                                                            reduce_all,
diff --git a/paddle/phi/kernels/impl/reduce_prod_grad_kernel_impl.h b/paddle/phi/kernels/impl/reduce_prod_grad_kernel_impl.h
index 6b93e98cec0168ab55e15e3401a72738f79d3a07..fb361e34205582bd3478696ade10602240b7d66b 100644
--- a/paddle/phi/kernels/impl/reduce_prod_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/reduce_prod_grad_kernel_impl.h
@@ -24,8 +24,8 @@ namespace phi {
 template <typename T, typename Context>
 void ReduceProdGradKernel(const Context& dev_ctx,
                           const DenseTensor& x,
-                          const DenseTensor& out_grad,
                           const DenseTensor& out,
+                          const DenseTensor& out_grad,
                           const std::vector<int64_t>& dims,
                           bool keep_dim,
                           bool reduce_all,
@@ -34,8 +34,8 @@ void ReduceProdGradKernel(const Context& dev_ctx,
                           DenseTensor* x_grad) {
   ReduceGradKernel<Context, T, funcs::ProdGradFunctor>(dev_ctx,
                                                        x,
-                                                       out_grad,
                                                        out,
+                                                       out_grad,
                                                        dims,
                                                        keep_dim,
                                                        reduce_all,
diff --git a/paddle/phi/kernels/reduce_grad_kernel.h b/paddle/phi/kernels/reduce_grad_kernel.h
index ee6f3d19a094d29546e82e7138933eceb96459d0..a4b472c445888fd45ff497884d309778b7447151 100644
--- a/paddle/phi/kernels/reduce_grad_kernel.h
+++ b/paddle/phi/kernels/reduce_grad_kernel.h
@@ -43,8 +43,8 @@ void ReduceMeanGradKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void ReduceProdGradKernel(const Context& dev_ctx,
                           const DenseTensor& x,
-                          const DenseTensor& out_grad,
                           const DenseTensor& out,
+                          const DenseTensor& out_grad,
                           const std::vector<int64_t>& dims,
                           bool keep_dim,
                           bool reduce_all,
@@ -55,8 +55,8 @@ void ReduceProdGradKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void ReduceMaxGradKernel(const Context& dev_ctx,
                          const DenseTensor& x,
-                         const DenseTensor& out_grad,
                          const DenseTensor& out,
+                         const DenseTensor& out_grad,
                          const std::vector<int64_t>& dims,
                          bool keep_dim,
                          bool reduce_all,
@@ -67,8 +67,8 @@ void ReduceMaxGradKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void ReduceMinGradKernel(const Context& dev_ctx,
                          const DenseTensor& x,
-                         const DenseTensor& out_grad,
                          const DenseTensor& out,
+                         const DenseTensor& out_grad,
                          const std::vector<int64_t>& dims,
                          bool keep_dim,
                          bool reduce_all,
diff --git a/paddle/phi/ops/compat/reduce_sig.cc b/paddle/phi/ops/compat/reduce_sig.cc
index 4bca0523801c1a94f90197c93cc495c2c4f56eeb..273badee623816ee32d25b52f8951dd0f854d8a4 100644
--- a/paddle/phi/ops/compat/reduce_sig.cc
+++ b/paddle/phi/ops/compat/reduce_sig.cc
@@ -149,7 +149,7 @@ KernelSignature ReduceMaxGradOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
   return KernelSignature(
       "max_grad",
-      {"X", GradVarName("Out"), "Out"},
+      {"X", "Out", GradVarName("Out")},
       {"dim", "keep_dim", "reduce_all", "in_dtype", "out_dtype"},
       {GradVarName("X")});
 }
@@ -158,7 +158,7 @@ KernelSignature ReduceMinGradOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
   return KernelSignature(
       "min_grad",
-      {"X", GradVarName("Out"), "Out"},
+      {"X", "Out", GradVarName("Out")},
       {"dim", "keep_dim", "reduce_all", "in_dtype", "out_dtype"},
       {GradVarName("X")});
 }
@@ -167,7 +167,7 @@ KernelSignature ReduceProdGradOpArgumentMapping(
     const ArgumentMappingContext& ctx) {
   return KernelSignature(
       "prod_grad",
-      {"X", GradVarName("Out"), "Out"},
+      {"X", "Out", GradVarName("Out")},
       {"dim", "keep_dim", "reduce_all", "in_dtype", "out_dtype"},
       {GradVarName("X")});
 }